Impact of Data Augmentation Using SMOTE and GAN on Machine Learning Predictions for Vehicular Routing Datasets

This study addresses the challenge of imbalanced datasets in Vehicular Ad Hoc Networks (VANETs). We introduce a methodology comparing the Synthetic Minority Over-sampling Technique (SMOTE) and Conditional Tabular Generative Adversarial Networks (CTGANs) to balance our dataset, which is derived from complex simulations using OMNET++, SUMO, OpenStreetMaps, and Veins. This approach enabled the creation of synthetic data that accurately reflects essential routing metrics, significantly reducing the need for additional time-consuming simulations to increase the dataset. Through the employment of T-distributed Stochastic Neighbor Embedding (T-SNE), we assessed the quality of our augmented data and identified the most critical routing metrics using L1 regularization. These insights were crucial for optimizing our neural networks with the Hyperband algorithm. We compared these results with the neural network model from previous work, known as the Multimetric Predictive ANN-based Routing Protocol (MPANN), demonstrating improvements in the model’s performance even with a reduced number of metrics, as evidenced by gains in AUC, precision, F1-Score, and Cohen’s kappa. Moreover, our findings suggest a pathway for future implementation in urban vehicular network simulation environments.